Is 82% better than grep because it is better, or because the split flatters it?

Read the setup before the number. That figure is top-5 file accuracy on a frozen 202-issue SWE-bench Lite split that siftr was never tuned on, against BM25 at 52% and grep at 22%, and the README's own position is that if you already know the exact words, use grep because it is instant and free.

Can I trust it not to hide the code I actually needed?

That is the design constraint. read lists every section it left out, filter keeps the full output on disk, and nothing is ever deleted or rewritten from the agent's context, so a wrong answer costs one extra read rather than lost information. Every tool is also instructed to fall back to grep or a full read when nothing fits.

Why does a Jev tool need an OpenRouter key?

The setup step asks for one and the privacy note explains the path: each call sends file paths, definition names and relevant snippets to OpenRouter, which routes them to TypeSafe. That is the thing to check before pointing it at code you cannot send to a third-party API. Offline tests run without a key, so the benchmark scripts are the parts that need one.

What happens when one request is slow?

It gets replaced rather than waited on: any request that has not answered after 2s is duplicated and whichever finishes first wins. Search also runs in two passes, ranking every file from its name and definitions before zooming into the best 30 to point at exact lines, which is why a 4,000-file repo still returns in about two seconds.

How do I install it and connect it to my agent?

Run the install script (curl -LsSf .../install.sh | sh on macOS or Linux, or the PowerShell equivalent on Windows), which brings its own Python via uv. Then siftr setup once to paste an OpenRouter key, and siftr agents install to add it as an MCP server to Claude Code, Codex, Cursor, Claude Desktop, Gemini CLI, Windsurf, Kiro, opencode, pi and omp. It backs up each config first and is safe to re-run.

What are the four tools and which ones are proven?

search ranks the file an issue needs (right file in top 5 for 82%), read keeps the relevant parts of a file (92% of edited lines while cutting 59%), pick chooses from a long list (81% for about 550 test files) and filter trims long logs but is experimental, because grep won on the only labeled log dataset.

How fast and how expensive is a search?

search takes about two seconds on a 4,000-file repo (p90 2.9s) and costs 1 to 2 cents on a large repo; read and pick take under half a second and cost well under a cent. Pricing is $0.042 per million input tokens, and the offline tests run with no key.

CommunityJev Practices & Patterns

Siftr

Four read-only MCP tools that rank files, cut a file down to the relevant part and pick from long lists, measured on a frozen SWE-bench Lite split.

See highlights

Open github.com

Stars: 7
Language: Python
License: MIT
Published by: Community
Author: Bentlybro
Last verified: Oct 2026

Bentlybro/siftr

User repository on GitHub

View on GitHub

Fast, cheap judgment for AI coding agents: semantic search, focused reads and list picking in ~2s. CLI + MCP server on TypeSafe Jev. Benchmarked on SWE-bench.

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Find the right code without reading the whole codebase. siftr gives AI coding agents (and you) four read-only tools that answer "where is it?" and "which part matters?" in about a second, for about a cent — so agents spend their tokens on the work, not on searching. Measured on SWE-bench Lite — 300 real GitHub issues from 12 Python projects, where we know which file and lines the real fix changed. Numbers are from a frozen test split of 202 issues that siftr was never tuned on.

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Routing Jev for Evaluation Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

search puts the right file in the top 5 for 82% of issues, against 52% for BM25 and 22% for grep, on 202 issues it was never tuned on.
read keeps 92% of edited lines while cutting 59% of the file, where a simple cut-off keeps 67% and cuts 75%.
A few dozen parallel requests replace checking files one at a time, and an unanswered request after 2s gets a duplicate with the first finisher winning.
siftr only suggests: read lists every section it omitted, filter leaves the full output on disk, and a wrong answer costs a read.

Watch out

The 300-issue benchmark was cut to a frozen 202-issue split, and filter lost to grep on the only labelled log dataset available, so it is marked experimental rather than claimed. The README's own advice is that if you already know the exact words, grep is instant and free. Each call sends file paths, definition names and relevant snippets to OpenRouter, which routes them to TypeSafe, so it is not usable on code you cannot send to a third-party API. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

daf-jev

Python client and decision toolkit for the Jev System One endpoint with noul, choice, and score primitives, batched asks, composition helpers, concurrent evaluation, and an MCP server.

Jev Practices & Patterns#community#jevbest#catalog-import

Community

docxology/daf-jevPython

jev-guardrails

Mock carrier support agent with the guardrail layer built twice behind one interface: chat-model JSON judging versus TypeSafe Jev typed questions over the same 25 rules.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

deepansh-saxena/jev-guardrailsPython

jev-mode

Python library and CLI that moves bulk semantic judgments — ticket triage, stale-file tagging, bug or question routing — out of the agent's context into typed Jev questions, with a measured comparison against an in-context control.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

ddfeyes/jev-modePython

jev-routing-experiment

Runs Jev as an LLM router over two public benchmarks, with a no-Jev ablation and RouterArena's own scorer so the numbers sit next to the public leaderboard.

Jev Practices & Patterns#community#jevbest#catalog-import

Community

TokenTrim/jev-routing-experimentPython

jev-benchmark

Runnable benchmark putting Jev, jev-preview and Claude Sonnet 5 through one real agent task — classifying a tool call as readonly, destructive, privileged or exfiltration — on the same 60 labelled cases, with raw per-call results committed.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

themsquared/jev-benchmarkPython

jev-filter

jev-filter is a CLI that filters tool output before it reaches your agent and, since 0.3, drives web pages and desktop apps too. Jev picks the next operation and its target among controls the program enumerates itself; the program re-checks and executes, and irreversible steps stop for confirmation.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

apixly-ai/jev-filterPython

From the Jev guides

Original write-ups on what this entry is for.

Designing the State: What to Include and What to Drop

Oct 12, 2026

Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.

Jev Practices & Patterns#state#design#primitives

Read article

Noul, End to End: One Probability and a Dozen Thresholds

Oct 12, 2026

A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.

Jev Practices & Patterns#noul#primitives#confidence

Read article

Scoring Without Measuring: Bands, Rubrics, and Composite Triage

Oct 12, 2026

A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.

Jev Practices & Patterns#score#primitives#scoring

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai